Difference-in-differences measures an effect without a randomised experiment. Take the change in an outcome for the group that received an intervention, subtract the change over the same period for a comparable group that did not, and the remainder is the estimated effect. Whatever moved both groups, such as a seasonal peak or a price change, cancels out.
It suits AI rollouts because most enterprise tools reach teams in waves anyway. Each wave can be compared with the teams still waiting, provided the order and dates are recorded. The largest field study of generative AI in customer support used this kind of staggered rollout to estimate a 15 per cent average productivity gain.
The method rests on one assumption: that the two groups would have moved in parallel without the change. Checking that the groups tracked each other before the rollout is the step most often skipped.